Schema: guard the destructive DSL rebuild vs the non-destructive `properties` edit

[Adron]

Problem

PUT /api/lists/{id}/schema has two body shapes with very different consequences:

• a schema DSL object → destructive: wipes and recreates all columns (and, per the docs, name becomes the list's new title);
• a properties array → non-destructive: rename a label, add/remove one column, row data preserved.

Shipping a single "Save schema" button that always sends the DSL form would silently destroy users' data.

Acceptance criteria

• Label edits, single add/remove, and reordering route through the properties shape.
• The DSL rebuild path is a separate, explicitly-labelled action.
• The rebuild requires in-place confirmation naming what is lost (no blocking native message box — see the dispatcher rule in CLAUDE.md).
• The rebuild path notes that schema.name overwrites the list title.
• Verified on a throwaway test list that the properties path preserves rows.

[Adron]

Live-verified findings from the schema service (#17 / PR #143) — three change the design here

⚠️ 1. The non-destructive properties form accepts only SIX types, not twelve

400 Unknown propertyType 'textarea'. Allowed: text, number, boolean, date, url, email

propertyType is also mandatory on every item (omitting it gives the same error for 'undefined'/'null').

Consequence: a list using textarea, datetime, tel, select, multiselect or priority cannot be edited non-destructively at all — every change to it has to go through the destructive DSL rebuild. That is a real, permanent constraint on #22's guard, not an edge case: the moment a user picks one of the six richer types in #18's builder, they lose safe editing on that list forever.

Exposed as ListFieldType.PropertiesEditable and ListSchema.SupportsPropertiesEdit so the UI can tell the user before they choose.

2. "Destructive" means columns, not rows — so #22's confirmation copy must not say "rows will be deleted"

Confirmed live. The DSL rebuild:

• drops and recreates every column with a new id, losing any validation / options / visibility the new DSL omits;
• leaves rows intact, but values whose keys no longer exist are orphaned in rowData — still stored, not shown, not validated.

And it renames the list: schema.name overwrote the title ('ZZ Throwaway Schema Probe' → 'ZZ Renamed By DSL Rebuild'), and schema.description overwrites the list description too — clearing it when omitted. #22's confirmation needs to say all three things.

3. Deleting a data-bearing column requires ?force=true

409 {…, "propertiesWithData": ["link"]}

?force=true then strips that key from every row. The shared EnsureSuccessAsync discards the propertiesWithData array, so a dedicated ListSchemaException carries Issues / PropertiesWithData / RequiresForce. #18 should name the affected columns in its confirmation rather than showing a bare 409.

4. The server never validates conditional visibility — #20 must

An unknown operator, a missing watched key, and a forward reference all return 200 and then silently never match. So there is no server-side safety net: ListSchema.Validate() does these checks client-side, and #20's editor should surface them rather than letting a rule quietly do nothing.

5. properties PUT definitively preserves row data — #22's guard is sound

A rename + column-add on a list holding two rows returned 200 and both rows came back with identical ids, identical version (still 1), and identical rowData. Column ids were preserved too, and the validationRules / helpText / visibilityCondition that the properties shape cannot even express survived untouched.

Note the envelope differs between the two directions: GET returns {"data":{…}}, the properties PUT returns {"properties":[…]} ordered by displayOrder — no data wrapper.

6. Schema-less lists are unaffected

Created without a schema they answer {"data":{"name":"<list title>","description":…,"fields":[]}} and accept rows with arbitrary, even nested, keys. Unknown keys are accepted on a schema'd list too — nothing is stripped. fields: [] cannot be PUT back (DSL must have at least one field), so ListSchema.Validate() catches that locally.

Also relevant to #21

Row-write validation returns 422 with details:[{field,message}], which the current plumbing flattens — so #21 can't currently attach a message to the offending field without unpacking that itself.

Separately, #144/PR #146 fixed a crash on the row-read path (ListDataRow.ListId was required but absent) and modelled version, which is the hook for optimistic concurrency on row edits.

[Adron]

Implemented in #164 — work continues in the PR from here. "Save columns" now routes renames, adds, single removals and reordering through the non-destructive properties shape and states its plan before running; the DSL rebuild is a separate "Rebuild columns…" action behind an in-place confirmation (no native dialog) that names all of it: new column ids, rows KEPT with orphaned values for dropped keys, the title overwritten from an editable box, and the description overwritten — cleared if that box is blank.

Verified live on throwaway lists with the bodies the real draft view-model serializes: the safe path answered 409 propertiesWithData ["link"], then 200 with ?force=true, leaving both rows with identical ids and the list's title/description untouched; the rebuild returned 200, renamed the list, kept both rows, and returned a null description when the DSL omitted it.

Worth recording because it is easy to get backwards: removing a column via properties ?force=true DELETES that value from every row, while a rebuild KEEPS it as an orphan — which is why the two confirmations are worded differently and the rebuild copy never says rows are deleted.